The principle is simple. When a robot gets stuck on a task, it picks up the solution found by another robot in the group that has already met the same difficulty. The authors call the method “learning from collective experience” (CEL). It is meant to get past the limit of learning based only on movements demonstrated by humans.
The robots worked on five tasks, at home and in the factory: arranging a bouquet, folding clothes, charging a phone, assembling a stool and packing a smartphone.
Across 600 real-world trials spread over the five tasks, the robots trained with this method succeeded in 97.5% of trials. On each task, their success rate beat the average of six human operators who piloted the robots remotely.
The larger the group, the better the robots perform. On smartphone packing, the hardest task, and with equal training time, the success rate rises from 86.7% to 98.3% when the group grows from 16 to 96 robots.
Update of 4 November: in its own presentation, AgiBot gives an example for phone packing: one robot completed 120 packing runs in a row with only 2 failures. According to the manufacturer, its success rate rose from 79.2% to 98.3%, and the number of failures fell from 25 to 2, or 92% fewer. The manufacturer says the group tolerates breakdowns. During the experiment with 96 robots, there were 9,834 task sequences and 30 robot breakdowns. Time lost to these breakdowns accounts for only 1.3% of total operating time.
Sources
- 로봇신문 Robot Newspaper · irobotnews.com
- 智元 AgiBot (新闻资讯) · agibot.com.cn
Also worth reading
Artificial intelligence
Rhoda AI wants robots to learn from web video, not just teleoperation
Illustration, not a photo of the robot